An online duplicate of a hidden bug

Preface

A previous project has been online for a long time , Recently, there was a sudden explosion Bug, Finally, the scope of impact will be evaluated Bug Upgraded to failure , Just because the amount of data affected is 10000 Bar or so , It has a certain impact on the business side .

But because it does not involve financial losses ,Bug Repair the data after repair , So the final level is also low .

Today I share with you this hidden online Bug, It can also be used for reference in work projects ~

Demand background

The theme ： B & B check-in return visit questionnaire

describe ：

For customers staying in B & B , You need to send the return visit questionnaire of this B & B check-in to the customer on the same day or the next day after leaving the hotel , In order to collect customers' comments or suggestions on the check-in experience

explain ：

Because of the large amount of data , It's using Hive Inventory storage data and logical processing

Problem analysis

1. Business logic

t2 surface ： Storage ( Yesterday, + At this time today ) Check out customer list information

t3 surface ： Store the customer list information that has sent the follow-up questionnaire

appear Bug The core business logic of is as follows ：

The equivalent association field used is ：

t2.id = t3.primary_id

from t2 surface Filter out t3 surface Data in , By using t2 surface As the main table left join Sub table t3 surface

And then determine t3.primary_id is null, The description has not been sent .

2. Problems arise

t2.id Field type is ：string type

t3.primary_id Field type is ：bigint type

When two different types of fields are equivalently associated , When t3.primary_id The actual value is more than 16 The value of bit is in MapReduce The field type is changed from bigint Type conversion to double type , At this point, there is a lack of accuracy .

for example ：

#  This is just an example , Does not represent the actual true value 
t2.id = 1000110000000000
t3.primary_id = 10001100000000009

The above example ,t3 surface There is already primary_id = 10001100000000009 The record of sent a follow-up questionnaire

here t3.primary_id = 10001100000000009 The value of exceeds 16 position , Convert it to double Type , The value is truncated and then correlated equivalently

Will appear t3.primary_id = 1000110000000000( Will end with 9 Get rid of ) And again t2 surface Medium id relation

There are already primary_id = 10001100000000009 The record of , So I think this record should not send a follow-up questionnaire , Directly filtered out

In the actual business, this record is not in the record form of the sent return visit questionnaire , Only because of the lack of precision, the two records on the inappropriate association are truncated and wrongly associated

result ： Lead to customers who should send a follow-up questionnaire , Not sent

Solution

1.Bug solve

At that time, this appeared online Bug The first solution is ： Type the equivalent associated fields , All converted to string Type and then conduct equivalent Association

Use cast function Do type conversion , This can solve the problem of lack of accuracy

2. Missing data processing

The modified code logic at the equivalent connection , According to the time range of business data affected , Do it again , Find out that you should have sent a follow-up questionnaire, but so Bug And the unsent customer list , Send a follow-up questionnaire .

summary

Through the previous analysis , You can see the online Bug More hidden , If you are not clear about the implicit conversion rules , This problem is likely to occur .

Today I will share with you this implicit transformation Bug, I hope that we can learn from one or two when there are similar problems in future work projects .

Except for Bug Remedial measures after , What can we do in advance in future projects ？

1. Once there is equivalent Correlation , You need to consider whether the field types of associated fields are consistent ;
2. When defining tables , It is necessary to consider the data range of the possible maximum value in combination with specific business conditions ;
3. Don't be afraid of this problem , Without thinking , As long as the equivalence relation is used cast function Conduct string Conversion of type ;